Research And Realization Of Object Detection Based On Deep Learning

Object detection is one of the most important basic problems in computer vision.Its main purpose is to locate and detect specific targets from static images.Object detection technology combines image processing,semantic segmentation,intelligent scene and automatic control technology,and has a wide range of applications in such areas as automatic driving,medical imaging,human-computer interaction,motion tracking and so on.Object detection algorithms based on traditional methods separate feature extraction and classification decision,and use manual extraction or design feature,which is difficult to get the ideal effect in the face of complex scenes.Since the concept of deep learning was proposed by professor Hinton,researchers have found that deep learning methods have great performance and speed advantages.Therefore,more and more researchers have tried to apply deep learning to various fields,including object detection,and proposed different models based on this model.Compared with shallow learning,deep learning is different in the following aspects: 1.It deepens the depth of network model and forms more hidden nodes;2.after the feature transformation of each layer,the images originally input to the network are transformed from one quantization space to another new quantization space,highlighting the characteristics of each layer and making subsequent classification and detection easier.The advantage of this method lies in its strong ability to express the model and highlight the characteristics of the target in surrounding non-targets.At the same time,this method also has a certain biological basis.This paper analyzes the basic models and methods of deep learning.From the perspective of computational complexity,computational efficiency and feature extraction ability,the traditional algorithm and deep learning model algorithm are deeply studied.On this basis,this paper proposes an improved object detection algorithm for multi-layer convolution feature aggregation and regression algorithm.Firstly,this paper uses the multi-level feature extraction model based on VGG-16 to obtain the object feature information.Different from the classical algorithm,Faster R-CNN only obtains the feature information from the last layer.This paper uses the output of the last three layers of VGG-16.The cascading functions are aggregated,so that the finally obtained feature map contains more lower layer information,which is beneficial to improve the accuracy of small object recognition.Then,the multidimensional feature map is reduced to a "thin" feature map through the deformable layer,and the convolution feature map from the feature extraction stage is processed by the regional extraction network RPN,and the images containing multiple suggestion boxes are output.Finally,the conditional probability model is used to calculate the boundary position of the target box to achieve more accurate target positioning,so as to improve the target detection accuracy.To verify the performance of the improved algorithm proposed in this paper,some experiments have been impletemented on the PASCAL VOC 2007/2012 dataset and MS COCO 2014 dataset.Besides,the performance of algorithm proposed by us also has been compared with that of classical network model based on deep learning owning the best accuracy at present.The experimental results show that the algorithm proposed in this paper achieves the best results in multiple datasets,and improves the average detection accuracy by 2.1 points.Compared with the classical object detection algorithm based on deep learning network,the algorithm proposed in this paper improves the positioning accuracy of object detection block through conditional probability algorithm.At the same time,a great deal of low-level feature information has been obtained by using the improved feature extraction model,so that the detection effect of small objects has been improved.